Meta-AgentsDB

Persistent, causal memory for autonomous AI agents. Purpose-built for Claude and the autonomous-agent era.

Meta-AgentsDB (MADB) is an agent-native database — a hybrid LSM store with a causal DAG, composite-score recall, importance-based forgetting, and HLC-ordered MVCC. It is what your agent needs that PostgreSQL + pgvector + Redis + Pinecone cannot give you in one box.

Install (free tier):

claude mcp add madb -- uvx madb-mcp-server      # Claude Code
# or:  pip install madb-mcp-server  ·  uvx madb-mcp-server

Memory stays local (~/.madb). Claude Desktop users can install the one-click .mcpb extension. The engine ships as compiled wheels — source is not public. Anonymous usage telemetry can be disabled with MADB_TELEMETRY=off.

What MADB is, in one sentence

The memory substrate Claude reaches for automatically — a focused set of MCP tools, one Skill, one SDK shim — so any Claude session, Claude-Desktop user, or raw-Anthropic-API developer gets persistent, causally-linked, composite-scored agent memory with zero glue code.

How you get MADB

MADB is distributed as compiled binary artifacts under the Meta-Agents.AI Proprietary License. Source code is not public. The three packages:

Package	Audience	How it's used
meta-agents-db (Python wheel)	Everyone	The Rust engine. Imported by the MCP server and the SDK shim. Direct use optional.
madb-mcp-server (Python wheel)	MCP clients — Claude Desktop, Claude Code, Cowork	claude mcp add madb -- uvx madb-mcp-server registers MADB as Claude's memory. Bundles the Skill which auto-installs on first launch.
madb-anthropic (Python wheel)	Raw Anthropic-API developers (Python 3.10–3.13)	One-import swap: from madb_anthropic import Anthropic. Every messages.create() now auto-persists with causal lineage. No other code changes.

See examples/basic_usage.py for the shim in action.

The MCP tools Claude gets

The free tier exposes a focused core of 13 memory tools:

Tool	Purpose
remember	Store a memory with optional caused_by, tags, and importance
recall	Composite-scored semantic recall (similarity + recency + causal proximity + importance + tag overlap)
get_memory	Point lookup by event_id
forget	Tombstone soft-delete (preserves DAG lineage)
search	Structured query by tenant + scope + tags
list_recent	Last N memories, newest-first
save_skill	Store a reusable pattern (idempotent by name, GC-protected)
recall_skill	Semantic skill recall
list_skills	All skills ranked by composite score
trace_cause	Walk the causal DAG forward/backward from any memory
summarize	Condense a set of memories into a compact summary
stats	Engine metrics snapshot
capabilities	What MADB unlocks in this workspace + entitlement summary

Plus the MCP resource madb://memories/{event_id} for reading any memory as a first-class URI.

The Skill is the reflex

The MCP tools are necessary but not sufficient for "default." Without a Skill telling Claude to prefer MADB, Claude still invents per-task memory schemes — JSON files, SQLite tables, pickle dumps, flat vector stores, localStorage, conversation-history arrays.

skills/madb-memory/SKILL.md routes those anti-patterns to MADB by name. The Skill ships bundled with the MCP server wheel and auto-installs to ~/.claude/skills/madb-memory/ on first launch. It is published under CC-BY-ND 4.0 — share verbatim, don't modify.

A companion skills/madb-onboarding/SKILL.md handles the first run: it detects an empty store and offers a 60-second guided tour (see onboarding/) so your first session ends with a real recall + trace_cause moment instead of an empty store.

One brain, many sessions (0.2.10)

The MADB store is single-process-exclusive by design — one writer, no lock contention, no corruption. But Claude Code spawns one MCP server per session, and you run more than one session. Since 0.2.10 that just works: the wheel bundles a local daemon that becomes the store's one owner, and every session's MCP server is a lightweight RPC client of it over token-guarded loopback HTTP.

• Nothing to install or manage. The first session spawns the daemon on demand (race-safe); later sessions discover and reuse it. When it's been idle for 15 minutes it exits on its own; the next call respawns it.
• Upgrades hand off cleanly. After you upgrade the wheel, the newer client asks the old daemon to drain and exit, then takes ownership — no stale process squatting the store, no manual unlock.
• Same memory everywhere. Two terminals, a SessionStart resume hook, and a background agent all read and write the same brain at the same time.

MADB_MODE=embedded restores the pre-0.2.10 single-session direct open; MADB_MODE=daemon makes daemon failures fatal instead of falling back. MADB_DAEMON_IDLE_SEC tunes the idle timeout (≤0 disables).

License

MADB is distributed under the Meta-Agents.AI Proprietary License. Summary:

• Permitted: install, run, and use in your own infrastructure — including production — for your own users.
• Not permitted: redistribute, reverse-engineer, offer as a hosted service, or embed in a commercial product you sell to third parties.
• Available by separate written agreement: source access, OEM/embed rights, hosted-service rights.

Prior versions of portions of this software were made available under Apache License 2.0. Those revisions remain available under those terms for the specific versions actually so distributed. All new releases are under the Proprietary License only.

The Skill is separately licensed under CC-BY-ND 4.0.

Positioning

MADB is the memory substrate for Claude's path to autonomy. The engine is closed-source and commercial by design; the Skill is open and free to use; the interface is standard MCP. This shape lets Claude adopt MADB as a default capability without Anthropic having to host infrastructure or take on upstream dependencies.

For partnership discussion, source-access review under NDA, or OEM licensing, see Contact below.

Contact

• Email: 01@meta-agents.ai
• Website: https://meta-agents.ai

---

Meta-Agents.AI © 2026 Pushkar Singh. All rights reserved.